AI and Copyright: The Scandal Was Never the Training

"Use every means necessary to take to court the artificial intelligence companies that plunder our work."
— From the resolution filed by members of the SACD, spring 2026

In the spring of 2026, three things are happening at once, and almost nobody is looking at them together.

Members of the SACD (the Société des auteurs et compositeurs dramatiques, the French collecting society that manages rights for tens of thousands of theatre, film and television authors) are circulating a resolution. It asks the June 25 general assembly to launch lawsuits against the providers of generative artificial intelligence. To even be discussed, it must gather 4,000 votes — and, reading article 34 III of the bylaws, that first vote isn't about the substance at all: it's about whether the assembly may debate it. A vote on whether they have the right to vote.

Meanwhile, in Parliament, the bill known as "Darcos" — after senator Laure Darcos — was passed unanimously by the French Senate in early April 2026. It seeks to establish a presumption that AI providers exploit cultural content: in other words, to reverse the burden of proof, so that the company must show it did not use a work, rather than the author having to prove it did. Sent on to the National Assembly, it was left off the agenda on May 11. The minister in charge of AI, Anne Le Hénanff, calls it "technically impossible to apply" and a source of "devastating legal risk."

And in the background, there's Anthropic. The American company behind the Claude assistant agreed, in September 2025, to pay $1.5 billion to settle a copyright dispute. Among the roughly 500,000 works involved are those of Gaël Faye, Guillaume Musso, Amélie Nothomb, Joël Dicker and Albert Camus. Authors had until March 23, 2026 to check whether they were eligible for compensation.

Three fronts, one shared anger. And yet I think we're fighting the wrong battle.

What the American courts have already decided

The Anthropic case has become a slogan: "AI plundered the authors." True. But it's not quite what the judge said, and the nuance changes everything.

In June 2025, Judge William Alsup, in the federal court for the Northern District of California, issued a two-part decision that both camps quote while keeping only their preferred half.

Part one: training a model on lawfully acquired books is "fair use" — the American doctrine that allows certain uses without permission. Alsup goes far with his words, calling the technology "exceedingly transformative." The pro-AI camp waves this phrase around like a blanket absolution.

Part two, conveniently forgotten: downloading pirated copies from shadow libraries like LibGen or PiLiMi (illegal repositories of scanned books) to build a permanent stockpile is not fair use. On that precise point, Alsup refused to clear Anthropic and sent the matter to a trial on damages. The $1.5 billion settlement didn't fall from the sky: it lands right here, on the acquisition, not on the learning.

Read the decision and the line jumps out. The problem was never that AI learns from books. The problem is that Anthropic stole them to do it. This isn't a lawyer's subtlety. It's the difference between buying a book and slipping it under your coat — except the shelf here held millions of titles.

Three cases, not two

This is why the current debate irritates me, on both sides.

The anti-AI camp talks as if "training a model" were a single, uniform, always-guilty category. The pro-AI camp talks as if "transformative use" settled everything, acquisition included. Both dodge the only question that matters: did the work enter the model with or without its author's consent?

There are three cases, not two.

The first: training without consent and without payment — the books vacuumed up from LibGen. That's a scandal, and authors are right to sue. No leniency there.

The second: training with agreement and payment. In November 2024, HarperCollins — one of the five big American publishers — signed a deal with Microsoft to license its nonfiction backlist: $5,000 per title, split fifty-fifty with the author, over three years, with mandatory opt-in and guardrails (the model's outputs are capped at 200 consecutive words or 5% of the book). An author who signs that knowingly isn't being robbed: they're selling a license. You can find the price indecent — I'll come back to that — but it isn't theft.

The third: training forbidden by default. That's the choice made today, out of caution, by online writing apps such as WriteControl, Extypis and Scribbook: don't feed any model with their users' texts, full stop. A perfectly valid third path, one that asks no one's permission.

To blur these three cases is to blur theft, sale and abstention. The Darcos presumption, by treating every AI provider as a presumed exploiter, lumps together the one who paid HarperCollins and the one who siphoned LibGen. It punishes the seller and the thief alike. I understand the intent — sparing the author a proof that's impossible to produce. But a presumption that doesn't distinguish lawful acquisition from plunder doesn't defend copyright: it dissolves the very line that makes copyright defensible. It's probably no accident that even the minister handling the file finds it unworkable: a text that presumes everyone guilty soon collides with reality.

Europe already picked its criterion — and it's consent

It gets lost in the noise of the American trial: in Europe, the question isn't framed as "fair use," which doesn't exist here. It's framed as opt-out.

The 2019 European copyright directive (known as the "DSM" directive, for Digital Single Market) created a text-and-data-mining exception — the famous "TDM," the automated analysis of large volumes of works. That exception permits training, unless the rightsholder has expressly objected in a machine-readable way. Translation: European law already decided, and it chose consent as the pivot. Not a total ban, not free licensing: a mechanism where the author can say no.

The Darcos presumption sits on top of that regime. Its real value isn't to ban training — it's to make the opt-out enforceable, to give the author a weapon when a company claims to have used nothing. Seen that way, the right fight isn't "let's sue every AI," it's "let's make refusal effective and proof reachable." A distinction that's anything but academic: it decides whom you go after, and for what.

The price is another fight — and a legitimate one

Then there's the question that stings, even inside the consent camp: $3,000 for a book — is that serious? American writers have already filed suit in California, targeting not just training practices but the settlement mechanisms themselves, which they consider tailored for the companies. They're not wrong. Three thousand dollars for a work that took three years to write, set against models valued in the tens of billions, looks less like compensation than a wholesale rate.

But — and this is the whole point — an unfair price is a separate problem from the principle. A license being badly paid doesn't make it illegitimate; it makes the negotiation bad. To conflate the two is to hand the pro-AI camp a rhetorical gift: all it has to do is rebut "$3,000 is too little" to look like it has rebutted everything else. The right answer isn't "no license," it's "a better license." The fight over the amount is fair. It must not become an alibi for refusing the principle of a market itself.

The same confusion, on the writing side

This confusion — mixing up cases that have nothing to do with each other — shows up identically on the neighboring question: "is AI going to write the books for us?"

Developers adopted AI in two years, and no one says they stopped being developers. They argue with the tool, reject most of what it offers, decide at every line. The problem doesn't start when an author asks for a synonym or has a stubborn paragraph reworded. It starts when 280 pages come out of a prompt and someone signs the cover — what I and others call vibe writing, a tracing of the "vibe coding" that serious developers already look down on. Assistance on one side, substitution on the other. The same line as for training: consent and involvement, not the tool itself.

Two debates, one disease: we file under a single word — "training," "AI that writes" — realities that are opposites. And as long as we conflate them, we fight badly.

The question to ask

I'm not saying the SACD is wrong to want to sue. On the pirates, it's right, and a little fighting spirit does no harm. I'll only note, without malice, that a collecting society scolding the models for "aggregating without transparency" is itself, like all its peers, regularly pressed by its own members to make its redistribution to the smallest authors more legible. Transparency is a good fight. It also applies to those who wage it.

The real question isn't "should we sue AI." It's: did this model steal, or did it license? As long as the law, the authors' societies and the op-eds ask the first question instead of the second, they hand the plunderers the finest screen of all: confusion. The thief loves it when we talk about "training in general." It drowns his case in his neighbor's, the one who paid.

I'm for consented and paid training. I'm against stolen training. And I defend the right, for anyone who wants it, to forbid it all on their own turf. Three positions, not one — and it's by distinguishing them, rather than melting them into a single indignation, that we actually protect authors.

---

Sources: Judge William Alsup's ruling, Bartz v. Anthropic, June 23, 2025 (Goodwin, Authors Alliance, Norton Rose Fulbright); Anthropic's $1.5 billion settlement, September 2025, and author compensation, March 23, 2026 deadline (ActuaLitté, ADAGP); the Darcos bill, passed by the Senate in early April 2026 and left off the National Assembly's agenda on May 11, 2026 (French Senate, ADAGP, SNAC, Decideurs Magazine); the SACD members' resolution for the June 25, 2026 general assembly (ActuaLitté); the HarperCollins–Microsoft deal, November 2024, $5,000 per title (Authors Guild, Publishers Weekly, Jane Friedman); Directive (EU) 2019/790, text-and-data-mining exception and opt-out (art. 3 and 4).